Friday, May 31, 2013

Noah Smith and me

I bring a comment thread debate over here.


Me:
I claim DSGE models generally need one arbitrary constant per phenominon to explain...The number of exogenous unexplained random variables keeps growing...DSGE seems not to have pulled away with this paper.

Smith:
I'd say instead that it's not clear that DSGE has pulled away. A new parameter has been added to explain a previously ignored phenomenon. But that doesn't mean that the model will then fail to "pull away" by explaining more phenomena in the future without the addition of yet more parameters.

me:
hmm I can't resist arguing about grammar.  "DSGE seems not to *have* pulled away" is in some past tense.  I didn't make a claim about what "will then" happen in the future. Well uh tenses uh grammar.  I stand by my claim.

Me: Fitting is not forecasting. It is easy to be wise after the fact.

Smith: Yes, sure. BUT, keep in mind that pseudo-out-of-sample forecasting, while not the same thing as true out-of-sample forecasting, is also not the thing as fitting.

Me:
I think that pseudo out of sample forecasting can be substantially the same as fitting.

The ways this can happen are
A)one tries something find it doesn't fit pseudo out of sample then one tries something else. This is fitting.  the Computer isn't doing all of the work but the process is trying different things till one gets a good fit.

B) Pseudo explaining an ad hoc reduced from model which fits the data.  First fit.  Then figure out a model which implies the coefficients of the reduced form model. In this case, step 1 is look for dramatic extraordinary behavior of time series in 2008 (many are well know and were discussed a lot in 2008-9 especially including huge quality premia).
Step 2 toss one in a VAR along with the DSGE state from some sort of filter. This is fitting not forecasting
Step 3 motivate the variable as an indicator of an otherwise hard to observe rare shock and again filter so that the deep disturbance in the problematic period is basically that newly introduced shock.

I think we have a case B and that step 2 is here

http://libertystreeteconomics.newyorkfed.org/2012/04/forecasting-the-great-recession-dsge-vs-blue-chip.html

my thought from 4/16/12

"'The other difference between this model and SWπ is the use of observations on the Baa-ten-year Treasury rate spread, which captures distress in financial markets.'

is a red flag. How is this incorporated ??? I think it is an ad hoc add on based on regressing outcomes on DSGE forecasts and the spread."


8:45 PM


No comments:

Post a Comment